Learning Submesoscale Processes at the Sea Surface

I will describe several of our team’s previous and ongoing projects to examine dynamics and submesoscale processes through deep learning analysis of remote sensing datasets and ocean general circulation model (OGCM) outputs. By learning the fundamental patterns of sea surface temperature (SST) at scales of ~1 to 100km, we have surveyed the incidence and geographic distribution of features (e.g. fronts) and processes (e.g. upwelling) that manifest in SST data.  The resultant model contains the salient dynamical features traced by SST.  I will then discuss an algorithm trained on the ECCO LLC4320 OGCM to accurately predict and infill partially masked (i.e. cloudy) scenes of SST.  This technique qualitatively outperforms standard, linear approaches and demonstrates the OGCM successfully captures submesoscale processes imprinted on SST.   Lastly, I will detail our current program to identify and characterize density fronts, first in OGCM outputs and then by combining multi-platform remote-sensing datasets (SST, sea surface height, winds, and salinity).  In turn, we aim to infer vertical transport through the mixed layer and to estimate the upper ocean heat content.